Alignment of Building Footprints Using Quasi-Nadir Aerial Photography
In this paper, we consider the alignment problem of building outlines, provided by openly available sources, and high resolution aerial images. This problem can be transferred to that of matching images with different modalities. After studying related works, we propose to minimize a cost function penalizing both color and gradient discrepancies. Semantic context is extensively taken into account, and additional information, such as classification result, can be integrated. Pyramid-based coarse registration and median-filtering-based outlier suppression were implemented as pre- and post-processing modules, respectively. We performed extensive tests with three very different datasets and achieved encouraging results, which were very stable once application of pre- and post-processing took place.